Image21-MLX-8bit

Built with Qwen. An independent native MLX quantization of Qwen-Image-2.1 for Apple Silicon, by ixim / iximbox. Non-commercial research and evaluation only under the original Qwen Research License. This is not an official Qwen release.

The complete checkpoint is approximately 18.68 GiB. DiT attention/MLP and language-encoder attention/MLP linears use 8-bit affine weights, group size 64, with BF16 activations. The full vision tower, token embeddings, language head, norms and DiT input/output/timestep/modulation layers retain floating-point precision. The VAE retains its original FP32 weights. See the exact 476 quantized modules in conversion.json.

Supports text-to-image, native reference-image editing and RGBA transparency through the included scripts. The wrapper preserves the sampler's alpha channel, which the pinned upstream text-to-image convenience method otherwise slices away. The runtime also enables the upstream fixed-prefix KV cache for text-to-image. Components are loaded and released by phase to reduce unified-memory use. See scripts/mlx_pipeline.py and its retained MIT attribution.

Only the 128GB M4 Max has been tested at 1024px. With phase loading, the six text-to-image cases have nearly the same MLX active-allocation peak: BF16 up to 16.44 GiB, Q8 up to 16.20 GiB. Editing peaks at 18.68 vs 16.20 GiB. A separate, cold two-step phase diagnostic found that the unquantized FP32 VAE decoder peaks near 16.19 GiB in both precisions; unused MLX cache reached about 42 GiB on the tested 128GB Mac. Cached buffers can be reclaimed and are not a minimum-memory requirement, but the active peak alone cannot certify 32GB or 48GB devices. For planning without device-specific validation, 64GB+ is a conservative budget, not a tested minimum. Smaller Macs need a complete 1024px trial with system-memory and swap monitoring. See phase-memory-diagnostic.json. 2048px and 10-reference editing have not been benchmarked. Allow roughly 50GB free disk space for ordinary inference; rebuilding also needs the original source and baseline checkpoints.

Visual findings / 视觉检查

Assistant inspection of all seven paired cases found no obvious severe quality regression in this small sample. English and Chinese text are correct in both precisions; cup count/order and sweater recoloring succeed; real RGBA transparency is retained. The English poster changes visibly in typeface and lamp structure, and the dragon changes expression/details. Other pairs are closer. Both portraits are cropped more tightly than requested, and both transparency samples contain slight near-zero background alpha. This is an unblinded, single-seed visual check, not evidence of lossless quantization or statistical equivalence.

Download / 下载

On an Apple Silicon Mac with uv installed:

uvx --from huggingface-hub==1.33.0 hf download ixim/Image21-MLX-8bit --local-dir Image21-MLX-8bit
cd Image21-MLX-8bit

Reproducible inference

uv venv --python 3.13 .venv
uv pip install --python .venv/bin/python -r requirements.lock.txt
.venv/bin/python -m scripts.infer --model . \
  --prompt 'A natural portrait in soft window light' --output outputs/portrait.png

.venv/bin/python -m scripts.infer --model . --input input.png \
  --prompt 'Change only the blue sweater to a red sweater. Preserve the person and background.' \
  --source-seed 42 --seed 1000042 --output outputs/edit.png

Transparent generation prompt: This is an RGBA image with transparency. A cute cartoon dragon sticker. The image has alpha channel and the background is transparent.

Measurements / 实测

Apple M4 Max, 40 GPU cores, 128 GiB unified memory, macOS 26.6.2, MLX 0.32.2. Both models use the same pinned MLX runtime and internal SSD; 1024×1024, 40 steps, CFG=1, no VAE tiling, one full warm-up per process, seven cases, one seed per case. Sequential component loading is enabled in both models. Per-image time includes component loading, prompt encoding, denoising and VAE decoding; PNG writing is excluded. This is a desktop session with other applications open. BF16 ran before Q8; there were no repeated or interleaved trials to control order and thermal effects.

Case	BF16 seconds	8-bit seconds	BF16 peak GiB	8-bit peak GiB
portrait	440.91	568.86	16.41	16.19
english_text	464.32	566.88	16.43	16.20
chinese_text	474.14	567.43	16.44	16.20
composition	480.91	562.52	16.44	16.20
texture	487.64	445.07	16.43	16.20
rgba	485.45	557.30	16.42	16.20
edit	606.29	665.00	18.68	16.20

Six-case text-to-image mean / 六类文生图平均：BF16 7.87 min, Q8 9.08 min. Q8 generation time relative to BF16 / Q8 相对耗时：**+15.3%** in this run.

Peak figures measure MLX allocations, not minimum physical RAM. Full raw records, original RGBA samples and the visual review are in evaluation. These are informal measurements, not an official benchmark. The BF16 baseline is the same MLX implementation; cross-runtime CUDA equivalence is not claimed. One seed per case is insufficient to establish statistical quality equivalence. There is no claim of lossless quantization or a guaranteed speedup.

Provenance

• Source: Qwen/Qwen-Image-2.1@b3179ad355be050328e483a9dfdd9e60cd62adfa.
• Runtime: Blaizzy/mlx-vlm@95b01ccad2d9f65a9e87f6a87bd1c5df69626261.
• Native MLX affine packed weights; not CUDA bitsandbytes INT8, FP8, GGUF or an MFLUX checkpoint.
• All three components passed exact tensor round-trip verification.
• conversion.json, modifications, Notice, license, file hashes.

To rebuild, first download the original source snapshot at the revision above to source-bf16 (requires additional disk space), then run from this repository:

.venv/bin/python -m scripts.audit --source source-bf16 --manifest evaluation/source-files.json
.venv/bin/python -m scripts.convert --source source-bf16 --bits 8 --output rebuilt-8bit

The audit verifies the original source file hashes before conversion. Rebuilding uses the original floating-point weights, never a dequantized CUDA INT8 checkpoint.

To repeat the paired evaluation after the source audit, create a native baseline and run each precision in its own process. This is a lengthy research workflow and needs extra memory and disk space beyond ordinary inference:

.venv/bin/python -m scripts.convert --source source-bf16 --bits 16 --output rebuilt-bf16
.venv/bin/python -m scripts.benchmark --model rebuilt-bf16 --output artifacts/eval/bf16
.venv/bin/python -m scripts.benchmark --model . --output artifacts/eval/8bit --edit-input artifacts/eval/bf16/portrait-s42.png
.venv/bin/python -m scripts.report